Humyn Labs, a Bengaluru-based physical AI research lab focused on data and evaluation for robots, released the second edition of BRIDGE, an automatic speech recognition benchmark report, on September 17.

It evaluated 23 voice AI models across 23 languages using human-verified noisy conversational audio to assess how well speech systems support human-robot interaction.

The report said overlapping speech raised average error rates from 41.2% to 45.2%.

It also said dialect shifts hurt accuracy in Bengali and Spanish, while models failed in different ways, including word substitutions, omissions, and fabricated insertions.

Across five non-Indic languages, the report said ElevenLabs averaged a 5.8% error rate, compared with 24.6% for GPT-4o-mini-transcribe. ElevenLabs also supports multiple languages.

Humyn Labs, which said it would deploy US$20 million to scale a human data layer for physical AI and robotics, also published the benchmark on its website.

Public descriptions reviewed did not specify the full list of models tested, the full list of languages covered, or how noise conditions were generated.

The report said BRIDGE covered Indic languages, Latin American Spanish, Brazilian Portuguese, and Vietnamese.

🔗 Source: Humyn Labs